NeurIPS 2026 Papers

Flexible switching between reasoning trajectories (i.e., thoughts switching) has significantly enhanced the reasoning capabilities of Large Reasoning Models (LRMs). However, existing models often switch excessively yet fail to sustain promising reasoning thoughts---a phenomenon termed ''under-thinking''. While recent efforts suppress switching to mitigate this, such over-correction may discard valuable trajectories. To address this challenge, we propose Steady Thought (ST), a novel thought-level preference optimization framework. ST formalizes under-thinking as a preference issue at switching points, forcing single-path continuations from identified points to yield high-quality reasoning. Then, ST performs thought-level preference optimization by treating the newly generated response as preferred and the original one as dis-preferred. Experiments across multiple models and datasets show that ST effectively reduces token consumption while maintaining or even improving accuracy. It reduces output length by up to 63.8% while improving accuracy by up to 11.4%. Further analysis suggests that ST helps models acquire a generalizable ability to reduce redundant switching across domains and languages.

Large language models deployed as personalized assistants must reason over long, evolving interaction histories. However, in long-term dialogue reasoning, relevant evidence is scattered across sessions, preferences may be revised over time, and standard long-context training fails to address these challenges under data scarcity and prohibitive computational costs. We propose StateTree, a data-driven RL pseudo-task that constructs a challenging auxiliary task from scarce dialogues with verifiable ground truth. StateTree augments multi-session dialogues with a tree-structured path-tracing task: key-value records are embedded across sessions to form a binary tree. Solving the task requires the model to traverse from root to leaf by retrieving records across sessions and comparing timestamps to resolve branches, then recover the hidden target question among distractor leaves. We apply curriculum RL training progressively increasing tree depth and introduce a compositional variant whose edges carry step-level reasoning fragments, training the model to compose partial cues into coherent queries. Trained on 10K-token contexts, StateTree generalizes to 128K tokens without full-length RL costs and exhibits capabilities including cross-session retrieval, temporal reasoning, knowledge update, and compositional multi-hop reasoning. StateTree outperforms both SFT and RL-based baselines while preserving short-context general reasoning. StateTree-7B achieves gains up to +23.60% on LongMemEval (128k), and StateTree-14B reaches 59.00% accuracy on LongMemEval, surpassing QwenLong-L1-32B (45.20%).

Spreadsheets are widely used for business analysis, financial modeling, reporting, and decision-making. However, most existing spreadsheet benchmarks evaluate isolated operations such as single-formula generation or local cell edits, and therefore fail to capture end-to-end workflows in realistic business settings. We introduce SpreadsheetBench 2, a workflow-level benchmark for spreadsheet agents that covers three task categories: generation, debugging, and visualization. The benchmark is constructed from authentic business data, including financial reports and corporate filings, and is annotated and validated by domain experts. The benchmark contains 321 tasks; each instance averages 11.8 worksheets and requires 593.5 cell modifications, reflecting large multi-sheet workbooks with cross-sheet dependencies. We evaluate eight frontier large language models under a unified multi-turn agent scaffold, and additionally include several LLM-based spreadsheet products as complementary baselines. Results show that current systems remain far from reliable on real-world workflows: the best model achieves 34.89% overall task accuracy, and debugging accuracy is as low as 12.00%. Trajectory analysis and a failure taxonomy further indicate that insufficient spreadsheet inspection and incorrect target-cell selection are the dominant bottlenecks. Together, these findings position SpreadsheetBench 2 as a challenging testbed for advancing reliable spreadsheet automation.

Speech tokenizers serve as the critical interface between continuous speech signals and discrete representations, and are widely used in speech systems, e.g., Automatic Speech Recognition (ASR) and Large Audio-Language Models (LALMs). In this paper, we find that speech tokenizers are fragile at the semantic level: slight perturbations can disrupt their semantic encoding process and consequently cause severe degradation in downstream ASR and LALM performance. To study this vulnerability, we propose T-SemAttack, a transferable semantic attack based on a semantic encoder ensemble. By jointly perturbing the representation spaces of multiple semantic encoders, T-SemAttack effectively disrupts the semantic content of speech while preserving perceptual quality, thereby inducing error tokens. We further analyze token fragility and layer-wise representation drift in relation to cross-model transferability and disruption of LALM attention. Our analysis reveals a progressive amplification chain in which small waveform perturbations are magnified through the tokenizer pipeline and lead to semantic collapse, e.g., the attack causes a 99.7% token change in the S3 tokenizer. Building on these findings, we introduce ROSETok, a robust speech tokenizer that combines noisy training with robust semantic distillation to improve reconstruction fidelity and downstream-task robustness. Extensive experiments on multiple datasets, 9 open-source and black-box ASR systems, and 4 LALMs demonstrate that T-SemAttack achieves strong transferable attack performance, while the proposed Robust Speech Tokenizer exhibits robustness under high-fidelity reconstruction and downstream tasks. Our code and demo are available at https://t-semattack.github.io.

Multilingual large language models often produce inconsistent reasoning and answers across semantically equivalent prompts in different languages. Prior work suggests that intermediate representations can be relatively language-agnostic, but generation becomes increasingly language-specific as models commit to discrete output tokens. This is problematic because language-specific lexical choices can cause semantically equivalent reasoning paths to diverge across languages. This motivates searching for a cross-lingual alignment signal that is less tied to any single vocabulary item or script. We propose **SOLAR**, an auxiliary objective for supervised fine-tuning that aligns _soft-token_ representations across languages, using English as a pivot. Soft tokens are probability-weighted mixtures over the vocabulary embeddings, yielding continuous representations that can aggregate information from semantically related tokens across languages. We then align each non-English soft-token summary to its English counterpart in the shared embedding space. Across four multilingual reasoning benchmarks, SOLAR improves accuracy by up to +17.7 points over the base model and +3.8 over standard supervised fine-tuning, with the largest gains on low-resource languages. SOLAR also strengthens final-layer cross-lingual similarity and substantially reduces language-cluster separability, suggesting that aligning soft-token representations helps preserve shared semantic structure during multilingual reasoning.

We study Slowly Annealed Langevin Dynamics (SALD), a sampler for tracking a path of moving target distributions and approximating the terminal target through time slowdown. We establish non-asymptotic convergence guarantees via a KL differential inequality, showing that slowdown improves tracking through contraction of intermediate targets and the complexity of the path. Motivated by training-free guided generation with pretrained score-based generative models, we further introduce Velocity-Aware SALD (VA-SALD), which explicitly incorporates the underlying marginal distributions of the pretrained model and uses slowdown to correct the additional deviation induced by guidance. This yields a principled framework for training-free guided generation for diffusion-based and related generative model families, together with convergence guarantees that clarify the roles of intermediate functional inequalities and guidance bias.

Language models are increasingly being deployed as user simulators, but their memory is far more reliable than that of real users. To measure this gap, we run a series of classic memory experiments from psychology on both humans and language models. Across tasks, we find that out-of-the-box language models exhibit better memory than humans, even when prompted to imitate human behavior. We then show that better prompting strategies and the use of a compactor can cause language models to forget content in a more human-like way. Using these methods, we show preliminary evidence that language models with human-like memory constraints can function as more effective user simulators in a downstream education task. Finally, we release human reference data and benchmarks to support future work on simulating human memory with language models.

Current virtual staining approaches offer the potential for time- and cost-efficient biomarker quantification in cancer diagnostics and prognostics. However, patch-wise inference for gigapixel whole slide images (WSIs) fails to maintain spatial continuity, yielding artifacts that cause catastrophic mismatches with ground-truth images. Although pathology Vision Foundation Models (VFMs) offer rich representations, their self-attention causes varying global contexts to produce inconsistent embeddings for the same physical region. We formalize and validate this ``context contamination'' as a sheaf-theoretic problem where these embeddings form a presheaf that violates the gluing axiom. To address this, we propose SheafStain, a new approach that reinterprets VFM features as sheaf-like sections for spatially and biologically coherent virtual staining. Specifically, SheafStain integrates class and patch tokens into a Schr\"odinger Bridge framework as sheaf-like sections. While the class token anchors biological consistency, patch tokens form a per-position spatial map. A backbone co-pretrained on Hematoxylin \& Eosin (H\&E) and Immunohistochemistry (IHC) yields non-degenerate cross-stain stalks, so a single VFM feature space supervises both input conditioning and output stain alignment. Departing from prior work that evaluates on isolated $256 \times 256$ patches and either random-crops or resizes the $1024 \times 1024$ ground truth, we translate at $256 \times 256$ and evaluate on the stitched $1024 \times 1024$ outputs across HER2, ER, PR, and Ki-67. SheafStain demonstrates promising results against six prior methods while mitigating patch-boundary stitching artifacts. Code will soon be released upon acceptance.

Expressive performers resonate with audiences because they transform sound into motion: rhythm invites steps, prosody evokes gestures, and energy shapes the flow of the body. However, existing audio-driven character systems often reduce audio to a task-specific condition, separately optimizing for co-speech gestures or music-driven dance. This fragmented view overlooks audio as the shared perceptual driver of performance, making it difficult to generate coherent full-body motion across speaking, dancing, and their intermediate states. We present VOXPERFORMER, a unified audio-to-motion framework that moves toward a generalized performance foundation model. Our key insight is that audio is not merely an external condition, but the hidden soul of performance style, governing how motion emerges, evolves, and transitions. To realize this, VOXPERFORMER introduces a structured motion prior to organize the full-body performance space, audio-to-prior distillation to transform sound into latent motion cues, a memory-augmented prior to resolve audio-motion ambiguity, and a decoupled diffusion generator with implicit transition matching for efficient and coherent synthesis. Experiments on BEAT2 and FineDance show that VOXPERFORMER remains competitive on speech-driven gesture generation while substantially improving hand-aware music-driven dance generation. User studies further show that participants prefer VOXPERFORMER, indicating stronger audio-motion correspondence and more natural perceptual resonance.

Recent progress in speech-driven 3D facial animation has improved vertex level reconstruction quality, but speech-consistent visible articulation remains difficult. This is because speech production follows structured and constrained lip jaw coordination and the mapping from acoustics to motion is inherently one to many. Motivated by the structured patterns of visible articulation, we propose a novel articulation-aware framework that models visible speech through directional lip motions and composes them into geometry consistent facial deformation. To represent visible articulation with three canonical directional motions, spreading, opening, and protrusion, we propose a Speech Articulatory Memory (SAM) that captures the correspondence between speech and directional articulatory motions under phonetic context through key-value memory structure based retrieval and decoding. Then, a Topology-aware Articulatory Composition (TAC) integrates the predicted directional motions under mesh topology to produce coherent 3D facial motion. Experiments on VOCASET and TFHP show that our method achieves the-state-of-the-art performance on standard reconstruction metrics and improves visible articulatory distance and velocity fidelity for key lip factors, while a user study confirms clear preference in lip sync and realism.

Large language models trained on unrestricted internet corpora inevitably embed information from the future, introducing lookahead bias that compromises the validity of backtests and causal inference in finance and the social sciences. Point-in-time language models—trained exclusively on text available up to each calendar date—eliminate this leakage by construction, but existing efforts typically produce models that lag substantially behind their unconstrained counterparts. We show that this performance gap can be narrowed through scale. Training decoder-only transformers with up to 4 billion parameters on 1 trillion chronologically filtered tokens from FineWeb, we construct a sequence of yearly model checkpoints spanning 2013–2024. Across a range of common-sense reasoning and language understanding benchmarks, our models approach the performance of leading open-weight models of comparable size (such as Gemma-3-4B and LLaMA-7B) trained on temporally unrestricted data, although a performance gap remains on several tasks. Finally, in a strict out-of-sample economic evaluation task, portfolios built from point-in-time embeddings achieve robust positive Sharpe ratios and perform close to full-sample counterparts that violate temporal validity, indicating that chronologically consistent language models can extract economically meaningful signals without relying on look-ahead bias. We release the complete pipeline—including dataset construction, training infrastructure, and evaluation code—to enable reproducible point-in-time language modeling and to support research applications that require strict temporal validity. Code is available publicly on GitHub (https://github.com/DjoFE2021/Point-In-Time-LLM) and model checkpoints as well as training datasets can be accessed on Hugging Face (https://huggingface.co/datasets/Diamegs).

Preference-based reinforcement learning (PbRL) is the dominant framework for aligning AI systems to human preferences. However, evaluation protocols for such data were designed for text and have not been validated for speech. We present the first ICC-based, controlled cross-modal study of human and synthetic preference annotations, comparing text and audio evaluations of identical semantic content across 100 prompts. We show that achieving $\textit{good}$ agreement within either modality (ICC(2,$k$) $\approx$ .80) requires $\sim$9 raters. At the same time, modalities show marked differences in how people report preferences: audio raters exhibit narrower decision thresholds, reduced length bias, and more user-oriented evaluation criteria, with near-chance cross-modality agreement. We demonstrate that synthetic ratings can be used to effectively predict inter-rater agreement, thus serving as an early signal for stimulus selection and proxy for human annotations. Together, these findings argue that evaluation protocols for audio preference data require modality-specific design rather than direct adaptation from text.

Small language models (SLMs) offer compelling advantages in cost, latency, and adaptability, but have so far lagged behind larger models on long-horizon software engineering tasks such as SWE-bench, where they suffer from pervasive action looping and low resolution rates. We introduce SWE-Protégé, a post-training framework that reframes software repair as an expert–protégé collaboration problem. In SWE-Protégé, an SLM remains the sole decision-maker while learning to selectively seek guidance from a strong expert model, recognize stalled states, and follow through on expert feedback. Our approach combines supervised fine-tuning on expert-augmented trajectories with agentic reinforcement learning that explicitly discourages degenerative looping and unproductive expert collaboration. We lightly post-train Qwen2.5-Coder-7B-Instruct to achieve 42.4% Pass@1 on SWE-bench Verified, a +25.4% improvement over the prior SLM state of the art, while using expert assistance sparsely (≈ 4 calls per task and 11% of total tokens).

Automatic speech recognition (ASR) systems are usually judged by aggregate transcript accuracy, but real failures often hinge on meaning: for example, a wrong dose, address, account number, or date can matter more than many harmless word errors. We introduce SMASH, a framework for finding targeted semantic failures caused by sparse INT8 weight-bit perturbations, a dominant model data type for ASR edge devices, in quantized sequence-to-sequence ASR models. Given an audio input and a meaning-critical span,SMASH searches for small weight changes that replace the intended meaning while keeping the transcript readable, plausible, and close to the original output. Each accepted fault is a certificate that a specific semantic substitution is reachable under a bounded perturbation budget. Empirical study focuses on numeric and scalar substitutions across three ASR backbones: Whisper-small.en (WS), Whisper-large-v3 (WLv3), and SeamlessM4T-v2-large (S-M4T); and three corpora: an LLM-assisted controlled corpus (LACC), MultiMed, and LibriSpeech. On LACC at a 20-bit budget ($B$), SMASH-hybrid accepts numeric substitutions on 24/45 WS targets, 19/44 WLv3 targets, and 5/34 S-M4T targets; reachability is lower on real-speech datasets. In matched LACC budget sweeps tested up to $B=40$, WS reaches its ceiling already at $B=20$, while WLv3 and S-M4T continue to gain accepted targeted substitutions through $B=40$. Human validation supports the gate-based labels: on a 105-item shared validation set, independent human-majority labels and a four-judge large- language-model ensemble agree on the composite clean_success label, with $\kappa=1.00$. The faults are sparse: the pooled median is three INT8 bit flips, with medians of two flips on WS, eleven on WLv3, and twelve on S-M4T. Entity and negation pivots are evaluated as secondary evidence in the appendix. SMASH shifts ASR robustness evaluation from "how much did the transcript change?'' to "which meanings can be changed, how plausibly, and with how few bit flips?''

Systematic reviews -- which requires comprehensive evidence collection and synthesis from large document corpora in response to targeted research questions -- are foundational in finance, social sciences, and other technical fields. Manual construction of evidence tables is labor-intensive, and recent LLM-based assistants relying on embedding or keyword based search often fail to meet the coverage standards of systematic reviews. We introduce SLIDERS, a novel LLM-based methodology for systematic reviews, by automatically assembling evidence tables tailored to research questions. In addition to extracting structured data from documents, SLIDERS can extract full-text excerpts that serve as direct evidence or as provenance for structured data. Core to SLIDERS is an automated evidence reconciliation agent that writes code to analyze and reconcile extracted evidence, bringing together information fragmented across documents, resolving inconsistencies across excerpts, and synthesizing overlapping findings into a coherent evidence table. In addition, SLIDERS allows users to ask follow-up questions in natural language to further explore the assembled evidence. We evaluate SLIDERS on three systematic-review-style tasks over large document collections. SLIDERS outperforms the best-performing baseline across benchmarks, remains near 90\% accuracy across 6M-11M-token corpora. On two new follow-up analysis benchmarks \system{} can answer 77.9% and 58.3% followup questions accurately.

Queer vernacular is rarely studied in NLP, despite advancements in resources and evaluation for other sociolects and informal language. Because of this, NLP systems often process queer language incorrectly, e.g., they misclassify it as hate speech or generate negative responses. To address this problem, we propose \textsc{Slaying}, the first real-world dataset of English queer slang. \textsc{Slaying} is community-validated, and includes over 400 queer slang terms that pertain to more than 20 queer subcommunities. We argue that queer language data resources have great potential in NLP -- e.g., as components of large pretraining corpora and as the basis for benchmarks -- and can improve queer users' experience of NLP systems. We leverage \textsc{Slaying} for two novel findings in support of this argument: (i) For a number of language models, we show that they are unbiased towards the queer community, but at the same time unable to process its language, i.e., absence of representation bias does not entail the absence of linguistic bias. (ii) Model performance on queer slang varies across queer subcommunities; it is generally worse for slang pertaining to African-American and Latine communities. These findings are relevant for both the queer NLP and the broader ML communities. \textsc{Slaying} is available to the public, and open to future revisions and extensions. \textbf{Warning:} This paper contains profane and potentially offensive language.

Reconstructing speech from non-invasive brain signals offers a promising pathway for restoring communication in individuals who are cognitively intact but unable to speak. Existing EEG-to-speech approaches formulate this task as \emph{acoustic reconstruction}, optimizing waveform fidelity while ignoring whether the generated speech preserves high-level semantic content. In this work, we revisit this formulation and argue that EEG signals carry not only acoustic but also semantic information. We identify two key limitations of prior methods: (1) the neglect of spatial relationships between EEG electrodes, and (2) the failure to exploit the semantic structure of the N400 paradigm, where congruent and incongruent trials reflect distinct semantic processing. We propose \textbf{SENSE} (\textbf{S}emantic-\textbf{E}EG \textbf{N}eural \textbf{S}peech Synth\textbf{E}sis), which combines a graph-based EEG encoder over electrode geometry with EEG Semantic Conditioning (ESC), aligning EEG to a pretrained semantic space using only congruent trials. On the N400 dataset, SENSE consistently outperforms prior methods on both acoustic and semantic metrics, and model-internal channel attribution suggests distributed reliance on auditory, sensorimotor, and centro-parietal regions, consistent with known speech-perception neuroscience. In the unseen-subject setting, SENSE trained on only two subjects already surpasses the strongest baseline trained on all eighteen subjects in word error rate, and matches it on acoustic metrics with as few as eight subjects.

Solubility prediction is a standard benchmark in computational chemistry, yet multi-solvent models which reportedly approach the experimental-noise ceiling (i.e. the aleatoric limit) are not yet reliable enough to be deployed. We argue that this gap is partly artefactual: published benchmarks differ in curation policies, evaluate on count-weighted RMSE that hides failure on tail-heavy solvent distributions, and treat the widely cited 0.6–0.8 log S inter-laboratory figure as the aleatoric ceiling even though it reflects worst-case, not expected, disagreement. We introduce SC3, a multi-solvent solubility benchmark built on BIGSOLDB v2.1 with three contributions: (i) a reproducible curation pipeline yielding 101,535 measurements over 1,327 solutes and 206 solvents, with a recalibrated aleatoric floor of 0.106 log S—roughly 6× tighter than the conventional figure; (ii) nested Gold/Silver/Bronze consensus tiers with per-point σ\sigma σ, three leakage-checked splits, and a multi-solvent metric suite (PS-RMSE, Z-RMSE); and (iii) a 31-model benchmark across six families, whose best Bronze PS-RMSE sits at ≈5×εaleatoric\approx 5 \times \varepsilon_{\text{aleatoric}} ≈5×εaleatoric, and we observe this is a gap unclosed by any deep alternative tested. We perform three follow-on analyses: data scaling, transfer from quantum-chemistry solvation energies, and feature-level attribution, which demonstrates that calibrated per-point uncertainty is a reusable infrastructure for diagnosis beyond point prediction.

Extracting sentence embeddings from MoE-based large language models is a promising direction, as they provide greater model capacity than dense models at comparable computational cost. Existing works perform the standard forward propagation process to extract embeddings, overlooking a key stability requirement in the MoE encoding process: semantically similar texts should be encoded into similar representations. In this work, we identify two key stability-related phenomena in MoE models: (1) routing stability varies across layers, and (2) shared and routed experts exhibit different levels of stability. To address these issues, we propose SAME, a Stability-Aware MoE Embedding extraction framework that dynamically allocates activated experts across layers and adjusts expert outputs to mitigate component instability and improve embedding quality. Specifically, SAME assigns fewer activated experts to layers with less stable routing, where layer-wise routing stability is measured by the average overlap between the experts activated before and after injecting Gaussian noise into each layer’s router inputs. Additionally, SAME reduces the contributions of routed experts to alleviate the impact of their instability. Notably, our method is training-free, seamlessly integrates with existing approaches, and incurs no additional inference overhead. Experiments on semantic textual similarity benchmarks demonstrate that SAME consistently improves the quality of extracted embeddings across multiple MoE backbones.